轻量化卷积神经网络红外目标识别性能分析与FPGA实现

随着深度学习应用于计算机视觉，其数据量大、网络层结构复杂，在硬件部署中存在资源不足、延时高等成为关键问题，本文通过分析五种较有代表性轻量化网络的优缺点，提出一种将轻量化网络应用到红外目标检测领域的基于MobileNet的轻量化网络改进，并以FPGA为硬件载体实现。该网络使用Tanh激活函数替代原有激活函数并简化网络层数，以适应红外目标的特征提取，针对深度学习目标检测算法在硬件实现方面存在的数据量大，资源占用大，运算延时高等问题，采用FPGA进行硬件实现。实验表明，在Xilinx Zynq-7020 XA开发板上，设定时钟频率100 MHz,输入图像大小为640×512,改进后的MobileNet在保证原相同精度情况下实现5.1 ms每张图像。

Infrared target recognition analysis and FPGA implementation based on lightweight convolutional networks

With the application of deep learning in computer vision,its large amount of data,complex network layer structure,insufficient resources in hardware deployment and high delay have become key problems. This paper,by analyzing the advantages and disadvantages of five representative lightweight networks,proposes a lightweight network improvement based on MobileNet,which applies lightweight networks to infrared target detection field. FPGA is used as the hardware carrier. In this network,Tanh activation function is used to replace the original activation function and the number of network layers is simplified to adapt to the feature extraction of infrared targets. In view of the problems existing in the hardware implementation of deep learning target detection algorithm,such as large amount of data,large resource occupation and high calculation delay,FPGA is adopted for hardware implementation. The experiment shows that on Xilinx Zynq-7020 XA development board,the clock frequency is set to 100 MHz and the input image size is 640×512. The improved MobileNet can achieve each image of 5.1 ms with the same accuracy as the original one.